Robot policies are adding explicit foresight, geometry, and task memory

Several papers train Vision-Language-Action models, or VLAs, to carry a compact action-related signal inside the policy. InternVLA-A1.5 uses 50 foresight tokens that condition a frozen video generator during training, then removes the video branch at inference and outputs 50-step continuous action chunks. CAC-VLA predicts latent actions from image and language tokens, using them to condition the continuous action expert through gated cross-attention. GeoMoLa takes a more geometric route: it learns discrete motion codes by predicting future point-cloud changes, then uses those codes for 6-DoF manipulation.

The reported gains are strongest where the task requires generalization rather than only imitation. CAC-VLA reports 98.3% average success on LIBERO and 89.5% on LIBERO-Plus supervised fine-tuning. GeoMoLa reports 84.7% average success on single-view RLBench across 10 tasks and 166 variations, ahead of RVT2 at 80.4%. InternVLA-A1.5 claims the best overall results across six simulation benchmarks, although the available excerpt does not include exact benchmark tables.

Long tasks are being handled with explicit subtask state and bounded skill vocabularies. Cortex uses a high-level vision-language model to maintain text memory and emit subtasks for a lower-level VLA. Its planner is restricted to 32 canonical manipulation primitives, with templates and reachability-aware annotations so the low-level policy receives executable commands.

DSWAM uses a related split. Its default path is a World Action Model executor that predicts dual-arm action chunks. A vision-language planner is activated when coarse household commands need decomposition. In real folding tests under the matched DeMaVLA setup, DSWAM raises success from 92.5% to 96.3% and reduces average completion time from 2'18" to 1'44". Cortex reports 65% success on 14-step real-world chemistry tasks and 55% on 14-step washing tasks, where the cited end-to-end baselines score 0%.

Two papers target failures that appear after a trained policy leaves its original setup. CamVLA addresses camera movement. It predicts the end-effector action in the camera frame, estimates a 6-DoF hand-eye transform from one RGB image, and converts the action into the robot base frame. On RLBench unseen viewpoints, π0 improves from 33.2% to 51.4% mean success with CamVLA. In real Franka tests, π0 + CamVLA also keeps higher success at 5°, 10°, and 15° camera offsets.

PRISM addresses the lack of scene-matched demonstrations. From one target-scene image and an instruction, it builds varied simulated “digital cousin” scenes, plans trajectories, and trains policies on the generated data. In sim-to-sim tests with 400 trajectories per task, PRISM reaches 98.0% on LIBERO for “Put milk in basket” with π0.5, compared with 48.0% for X-Sim and 14.0% for RoboTwin 2.0.

Deform360 adds a data-heavy line of work to the same control problem. It collects 198 deformable objects across 1,980 interaction sequences with 41 synchronized cameras and tactile-equipped grippers. The dataset includes cloth-like, rope-like, and volumetric objects, with 13 manipulation primitive types across 17 categories.

The key contribution is the paired sensing. The pipeline reconstructs per-frame geometry, tracks up to 1,600 points per view, lifts tracks into 3D, and refines them with tactile contact consistency. The summary reports a Chamfer distance error of 2.71×10^-5 m² for visuotactile tracking, versus 1.41×10^-4 m² for visual-only tracking. Visual-to-contact prediction reaches 88.67% mean accuracy across 36 filtered views. This gives future robot world models a concrete benchmark for contact-rich deformable manipulation.